FACADE DEFECTS CLASSIFICATION FROM IMAGE DATASET USING DEEP LEARNING

Abstract

Defects inspection conducted for buildings is a compulsory work in reality. With the development of technology, it is necessary to find a method to increase the productivity and efficiency of inspection works. Artificial Intelligence (AI) has been very popular in the construction industry and Machine Learning (ML) is one element of it. ML consists of a subset called Deep Learning (DL) which has a strength of automatic classification through input data and learning complex functions. Thus, applying DL in facade defects classification is an excellent method and this method has to be based in 2D images for easy operation. Image data collection is done by taking images for facade defects of public housing in Singapore. In DL, the Convolutional Neural Network (CNN) is the most common and efficient algorithm for image applications. Moreover, the Visual Geometry Group with 16 layers (VGG-16) network which is a systemic CNN model as a pre-trained model and incorporates a transfer learning: Fine-Tuning method to develop the classification model to train 20,909 images and test 350 images per epoch. The model will output the accuracy and loss results, plot out the normalised confusion matrix and the receiver operating characteristics (ROC) curves for analysis.